check-citations

Verify academic citations against CrossRef, Semantic Scholar, and OpenAlex. Detects AI-hallucinated references, chimeric citations (real title + wrong authors), and suspicious patterns before submission.

When to Use

• After writing or editing a .bib file with AI assistance
• Before submitting a paper, thesis, or report
• When reviewing AI-generated literature sections
• As a CI/CD check in LaTeX manuscript pipelines
• When auditing existing bibliographies for dead or fabricated references

Background

• 6-55% of AI-generated citations are fabricated (varies by model/domain)
• 100+ hallucinated references found in NeurIPS 2025 accepted papers
• Universities increasingly treat fake citations as academic misconduct
• Three hallucination types: fully fabricated, chimeric (real title + wrong authors), modified real (slightly altered metadata)

Usage

Quick Check (Single File)

python scripts/citation_checker.py references.bib

Check All .bib Files in a Directory

python scripts/citation_checker.py path/to/report/

JSON Output (CI/CD Pipelines)

python scripts/citation_checker.py references.bib --json

Verbose Mode (Debug API Responses)

python scripts/citation_checker.py references.bib --verbose

How It Works

Cascading Multi-Source Verification

Each citation is checked against three independent databases:

Verification logic:

• Found in 2+ sources with matching title → verified (high confidence)
• Found in 1 source only → suspicious (manual check recommended)
• Found in 0 sources → not_found (likely hallucinated)

Chimeric Detection

When a citation's title matches a real paper but the authors don't overlap at all, it's flagged as a possible chimeric hallucination — the most dangerous type because the title looks real on Google Scholar.

Red Flag Heuristics

• Invalid DOI format (doesn't start with 10.xxxx/)
• Suspiciously generic title patterns ("A Comprehensive Survey of...")
• Future publication year
• Missing author or year fields
• Single-word author names (incomplete metadata)

Exit Codes

Dependencies

pip install requests

No API keys required — uses free tiers of all three databases.

Accuracy (Tested)

The core guarantee: fake papers are never marked as real.

Limitations

• Papers without DOI that have many derivatives (e.g., BERT without DOI) may not be found via title search alone — always include DOIs when available
• Semantic Scholar free tier rate-limits at ~100 requests/5 minutes — batch verification is slower
• Cannot verify papers behind paywalls or not indexed in any of the three databases
• Book chapters, technical reports, and grey literature have lower coverage

Integration with LaTeX Workflows

Pre-commit Hook

#!/bin/bash
# .git/hooks/pre-commit
python scripts/citation_checker.py references.bib --json > /tmp/cite_check.json
NOT_FOUND=$(python3 -c "import json; d=json.load(open('/tmp/cite_check.json')); print(d['summary']['not_found'])")
if [ "$NOT_FOUND" -gt "0" ]; then
    echo "BLOCKED: $NOT_FOUND unfound citations. Run 'python scripts/citation_checker.py references.bib --verbose' to investigate."
    exit 1
fi

GitHub Actions

- name: Check citations
  run: |
    pip install requests
    python scripts/citation_checker.py references.bib --json > citation_report.json
    python -c "
    import json, sys
    r = json.load(open('citation_report.json'))
    if r['summary']['not_found'] > 0:
        print(f'FAIL: {r[\"summary\"][\"not_found\"]} citations not found')
        sys.exit(1)
    print(f'PASS: {r[\"summary\"][\"verified\"]} verified, {r[\"summary\"][\"suspicious\"]} suspicious')
    "